T cell receptor sequencing of early-stage breast cancertumors identifies altered clonal structure of theT cell repertoireJohn F. Beausanga, Amanda J. Wheelerb, Natalie H. Chanb, Violet R. Hanftb, Frederick M. Dirbasb, Stefanie S. Jeffreyb,and Stephen R. Quakea,c,d,1

aDepartment of Bioengineering, Stanford University, Stanford, CA 94305; bDepartment of Surgery, Stanford University School of Medicine, Stanford, CA94305; cDepartment of Applied Physics, Stanford University, Stanford, CA 94305; and dChan Zuckerberg Biohub, San Francisco, CA 94518

Contributed by Stephen R. Quake, October 14, 2017 (sent for review August 7, 2017; reviewed by Ramy Arnaout and Curtis G. Callan, Jr.)

Tumor-infiltrating T cells play an important role in many cancers,and can improve prognosis and yield therapeutic targets. Wecharacterized T cells infiltrating both breast cancer tumors and thesurrounding normal breast tissue to identify T cells specific to each,as well as their abundance in peripheral blood. Using immuneprofiling of the T cell beta-chain repertoire in 16 patients withearly-stage breast cancer, we show that the clonal structure of thetumor is significantly different from adjacent breast tissue, with thetumor containing ∼2.5-fold greater density of T cells and higherclonality compared with normal breast. The clonal structure ofT cells in blood and normal breast is more similar than betweenblood and tumor, and could be used to distinguish tumor from nor-mal breast tissue in 14 of 16 patients. Many T cell sequences overlapbetween tissue and blood from the same patient, including ∼50% ofT cells between tumor and normal breast. Both tumor and normalbreast contain high-abundance “enriched” sequences that are absentor of low abundance in the other tissue. Many of these T cells areeither not detected or detected with very low frequency in theblood, suggesting the existence of separate compartments ofT cells in both tumor and normal breast. Enriched T cell sequencesare typically unique to each patient, but a subset is shared betweenmany different patients. We show that many of these are commonlygenerated sequences, and thus unlikely to play an important role inthe tumor microenvironment.

T cell receptor | breast cancer | repertoire sequencing

The immune system is thought to play an integral rolethroughout the life cycle of many cancers, including pre-

venting initiation, suppressing development, and influencingtreatment and patient outcomes (1, 2). Genomic alterations intumors create immunogenic targets that can be recognized asnonself and eliminated by cytotoxic CD8+ T cells (3). At thesame time, this process imposes a selective pressure on the tu-mor to evade this surveillance (4, 5), sometimes hijacking im-mune mechanisms that can then be targets for immunotherapy(6, 7). Breast cancer is less immunogenic than other cancers (8),but tumor-infiltrating lymphocytes (TILs) have been observed inall subtypes and shown to have prognostic value in human epi-dermal growth factor receptor 2 (HER2)-positive breast cancerand triple-negative (estrogen receptor-negative, progesteronereceptor-negative, and HER2 receptor-negative) breast cancer(recently reviewed in refs. 9–12).High-throughput DNA sequencing of the recombined V(D)J

region of the T cell receptor beta-chain (TCRB) has become astandard technique for quantifying the distribution of millions ofT cells in a biological sample (13–15). One cell in 100,000 is re-liably detected (16), improving clinical monitoring of pathogenicimmune cells in various blood cancers (17). Large (>600 individ-uals) public databases of TCRB data have been generated andused to infer individual major histocompatibility complex (MHC)alleles and cytomegalovirus (CMV) exposure (18). Even thoughrecent developments in single-cell methods can identify both alpha-

and beta-chain sequences (19–21), the peptide-MHC target ofeach T cell is generally not known. Regardless, characterizing theT cell receptor repertoire over a range of conditions can provideinsight into the subset of T cell sequences that may be relevant ina variety of clinical applications (22).Exploratory studies of the T cell repertoire in tumors from

several cancers have found differing repertoires between co-lorectal tumors and adjacent mucosal tissue (23), intratumoralheterogeneity in renal (24) and esophageal (25) carcinomas,spatial homogeneity in ovarian cancer (26), two subgroups ofT cell repertoires in pancreatic cancer (27), increased clonality ofCD4+ T cells in non-small cell lung cancer (NSCLC) comparedwith CD19+ B cell and CD8+ T cell compartments (28), andstereotyped shifts in the repertoire after cryoablation and im-munotherapy in breast cancer (29). Large-scale genomic studiesextracting T cell sequences from bulk RNA-sequencing datafrom The Cancer Genome Atlas (TCGA) observed a strongcorrelation between T cell diversity and tumor mutation load in avariety of cancer types (30). Exome data from TCGA were usedto derive an immune DNA signature for breast cancer thatcorrelated with clinical outcomes (31). Single-cell sequencing hasalso been applied to immune cells infiltrating breast cancer. Inone study (20), a subset of CD8+ T cells with matching alpha-and beta-chains was detected in breast tumors and sentinellymph nodes from multiple patients.In this study we use TCRB sequencing to determine the T cell

repertoire in matched samples of peripheral blood, tumor, andadjacent normal breast tissue from 16 patients with early-stage

Significance

The recent advances in cancer immunotherapy motivated us toinvestigate the clonal structure of the T cell receptor repertoirein breast tumors, normal breast, and blood in the same indi-viduals. We found quantitatively distinct clonal structures in allthree tissues, which enabled us to predict whether tissue isnormal or tumor solely by comparing the repertoire of thetissue with blood. T cell receptor sequences shared betweenpatients’ tumors are rare and, in general, do not appear to bespecific to the cancer.

Author contributions: J.F.B., S.S.J., and S.R.Q. designed research; J.F.B., A.J.W., N.H.C.,V.R.H., and F.M.D. performed research; J.F.B. contributed new reagents/analytic tools;J.F.B., S.S.J., and S.R.Q. analyzed data; and J.F.B., S.S.J., and S.R.Q. wrote the paper.

Reviewers: R.A., Beth Israel Deaconess Medical Center; and C.G.C., Princeton University.

The authors declare no conflict of interest.

This open access article is distributed under Creative Commons Attribution-NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND).

Data deposition: The data reported in this paper have been deposited in the immuneACCESSdatabase, https://clients.adaptivebiotech.com/pub/beausang-2017-pnas.1To whom correspondence should be addressed. Email: quake@stanford.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1713863114/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1713863114 PNAS | Published online November 14, 2017 | E10409–E10417

IMMUNOLO

GYAND

INFLAMMATION

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

breast cancer. We characterize T cell repertoires from thesetissues and show that the clonal structure of the tumor is dif-ferent from normal breast and blood, and can be used, in prin-ciple, to distinguish tumor from normal breast. We identify anddescribe a subset of “enriched” TCRB sequences with highabundance in each tumor and absent or low abundance in nor-mal breast. Lastly, we characterize a subset of complementarity-determining region 3 (CDR3) sequences that are shared betweendifferent patients and argue that most of these are likely to becommon sequences that are unlikely to play an important role inthe tumor microenvironment.

ResultsStudy Population. Of 16 invasive breast cancers, 12 tumors wereestrogen receptor-positive (ER+), progesterone receptor-positive(PR+), and HER2-negative (HER2−). Among these 12 tumors,eight were invasive ductal carcinomas and four were invasivelobular carcinomas, grades 1–2 with low proliferation rates (11 of12 had Ki67 < 15%) and 1 to 3 cm in size (10 of 12 were less than3 cm in diameter). The remaining four tumors had various receptorstatuses, including an 8-cm triple-negative (ER−/PR−/HER2−)high-grade tumor with a high proliferation rate (Ki67 of 50–70%)and an ER+/PR+/HER2+ invasive mucinous carcinoma. Two pa-tients received neoadjuvant chemotherapy before surgery, and onepatient had hepatitis C infection (Table 1).

Breast Tumors Contain a Larger Proportion of T Cells and a MoreClonal Repertoire than Normal Breast Tissue. V(D)J rearrange-ments within the TCRB locus are PCR-amplified with V- andJ-gene–specific primers, and the sequence of the CDR3 was de-termined using high-throughput sequencing of genomic DNA iso-lated from peripheral blood, tumor, and normal breast tissue (Fig.1A). The absolute abundance of input template molecules for eachnucleic acid sequence, which is an estimate of the number of inputT cells, is determined from synthetic spike-in control sequences(details are provided in Materials and Methods). The abundanceof individual T cell clonotypes is determined by combiningtemplates from molecules with the same TCRB V-gene family,J-gene segment, and productive (i.e., in-frame) CDR3 aminoacid sequence. A total of 2,651,842 templates across 1,097,674unique clonotypes were detected, with individual abundances

ranging from one to 28,508 templates per clonotype (details areprovided in SI Appendix, Table S1).The size of the TCRB repertoire is different for each tissue.

The number of productive templates represents the number ofT cells in each sample, and is greatest in blood (median of99,500 per sample), followed by tumor (median of 42,500 persample) and normal breast tissue (median of 17,200 per sample;Fig. 1B). Normalizing the number of productive templates by thenumber of input cells (SI Appendix, Table S2) results in an es-timate of the overall T cell density in each tissue. Tumors contain∼2.5-fold higher density of infiltrating T cells than adjacentnormal breast tissue (median of 9.6% and 3.8%, respectively;P < 0.0005; Fig. 1C), with both containing a lower density ofT cells than found in circulating peripheral blood mononuclearcells (PBMCs; median of ∼74%).The structure of the repertoires also differs between the three

tissues. The number of unique clonotypes is approximately two-fold larger in tumor (median of ∼12,500) compared with normalbreast (median of ∼7,000; P < 0.01), with fourfold more detectedin blood (median of ∼46,000) compared with tumor (SI Appendix,Fig. S1A and Table S1). The corresponding fraction of T cells withunique sequences, however, is lower in tumor (median of 0.31)than in either blood (median of 0.64; P < 0.0005) or normal breast(median of 0.49; P < 0.005; Fig. 1D). The cumulative abundancewithin the top 50 clonotypes is similar between tumor and normalbreast (median of ∼27%), with both having approximately three-fold higher abundance than blood (median of ∼9%; Fig. 1E). Theclonality in the tumor (median of 0.16) is higher than in blood(median of 0.10; P < 0.05) and normal breast tissue (median of0.12; P < 0.09; Fig. 1F). Both metrics indicate that the repertoireof T cells in the tumor is less diverse than in normal breast tissue.Recent tools to estimate total repertoire diversity (32) indicatethat the differences we observe in unique clonotypes and diversity(SI Appendix, Fig. S1 B and C, respectively) are robust to thesampling depth used here. Neither T cell density nor clonality dis-criminates between ductal (n = 8) and lobular (n = 4) ER+/PR+/HER2− breast tumors (SI Appendix, Fig. S2). The small study sizeprevents further tumor subtype analysis.

Abundant Clonotypes Are Often Detected in Multiple Tissues from theSame Patient. The overlap between two repertoires is definedhere as the fraction of T cells in one sample with at least oneidentical clonotype in the other sample, and is calculated over allpairs of samples (SI Appendix, Fig. S3). Averaging over theintrapatient tissue combinations shows that approximately half ofthe templates in tumor and normal breast overlap with eachother and with blood, whereas 25–30% of the templates in bloodoverlap with either tissue (Fig. 2A). Approximately 100-fold lessoverlap is observed between tissues from different patients (Fig.2B). Clonotypes with large abundance in one tissue are morefrequently detected in other tissues from the same patient (Fig.2C and SI Appendix, Fig. S4). The degree of this overlap is highlyvariable from patient to patient, but the trends are similar withinthe different tissue compartments of the same patient (SI Ap-pendix, Fig. S5). The average overlap across each patient’s tissuesis strongly correlated with the average clonality (Fig. 2D), in-dicating that the observed patient-to-patient variability is likely aproperty of the underlying repertoire. This is consistent withhigher clonality repertoires containing more abundant clono-types, which have a greater probability of being detected inmultiple tissues. Sequences detected in all three tissues are alsobiased toward high-abundance clonotypes, with the fraction oftemplates detected in all three tissues nearly 30% compared withless than 10% between any two tissues (SI Appendix, Fig. S6).

Tumor, Normal Breast, Blood, and Contralateral Breast ClonotypeAbundances in a Single Patient. Despite the significant fractionof overlapping T cells between tissues within the same patient,

Table 1. Clinical information for each patient

Patientidentification Age, y Type Size, cm Grade Ki67, % ER/PR/HER2

BR01 66 IDC 1.2 2 <5 +/+/−BR07 71 IDC 1.2 1 1–5 +/+/−BR13 58 IDC 1.0 2 10–15 +/+/−BR15 72 IDC 2.0 2 5–10 +/+/−BR16 56 IDC 1.8 2 5–10 +/+/−BR17* 66 IDC 2.1 2 10–15 +/+/−BR18 49 IDC 2.0 1–2 21 +/+/−BR21 45 IDC 2.5 1 1–5 +/+/−BR19 68 ILC 2.4 1 5–10 +/+/−BR20 67 ILC 4.7 1 10 +/+/−BR14 61 ILC 1.9 1 5–10 +/+/−BR26† 43 ILC 3.2 1 5–15 +/+/−BR22 36 IMC 2.0 2 <5 +/+/+BR05 54 IDC 2.5 2 10–15 +/−/−BR25 57 IDC 2.2 3 20–30 −/−/+BR24† 55 IDC 8.0 3 50–70 −/−/−

IDC, invasive ductal carcinoma; ILC, invasive lobular carcinoma; IMC, in-vasive mucinous carcinoma.*Indicates patient with hepatitis C infection.†Indicates patients receiving therapy before tumor removal.

E10410 | www.pnas.org/cgi/doi/10.1073/pnas.1713863114 Beausang et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

some highly abundant clonotypes in one tissue are either notdetected or detected at very low levels in the other tissues.Scatter plots of the abundance of each TCRB sequence betweentumor and normal breast for all patients (SI Appendix, Fig. S9)indicate that such clonotypes exist over a range of abundances,with varying amounts detected in blood from each patient (SIAppendix, Figs. S10 and S11).In patient BR21, for example, 1,802 sequences (comprising

∼45% of the templates) overlap between tumor and normalbreast, and span a broad range of abundances from 0.002 to 11%(Fig. 3A). One hundred fifty-four clonotypes are present withabundance greater than 0.1% in either tissue, including 32 in tu-mor (9.5% of templates) that are missing from normal breast and16 in normal breast (4% of templates) that are missing from tu-mor. To characterize this subset of “tumor-enriched” sequences,we define them to include sequences that have abundance greaterthan 0.1% in tumor and at least 32-fold greater abundance relativeto normal breast (details are provided in Materials and Methods).In the tumor, approximately half (42 of 81) of the abundant se-quences (29% of templates) are enriched, whereas in normalbreast, approximately one-third (32 of 89) of abundant sequences(9% of templates) are enriched. Unlike BR21, where the top eightclonotypes in the tumor are all enriched, most patients typicallyhave several clonotypes with high abundance in both tumor and

normal breast tissue (SI Appendix, Fig. S9), suggesting that high-abundance clonotypes in the tumor are not specific to the tumor.At the sequencing depth used in this study, ∼60% of all TCRB

templates in tumor and normal breast have clonotypes that aredetected in blood (∼24% of the templates in blood), including asimilar fraction of tumor-enriched (24 of 42; Fig. 3B) andnormal-enriched (17 of 32; Fig. 3C) clonotypes. Most tumor-enriched clonotypes correspond to low-abundance clonotypesin blood (median is less than 0.001%), which, together, total only0.15–0.17% of templates.BR21 is the only patient with a second normal sample from

the contralateral breast (Fig. 3D). The top clonotypes in the twonormal tissues are very similar, with 109 of 113 sequences (∼18%of templates) detected in both tissues. Unlike the case in blood,where the shared clonotypes have very low abundance, theshared clonotypes in the contralateral breast are highly corre-lated with those in normal breast tissue (Pearson correlation =0.60; SI Appendix, Fig. S8A). All of the normal-enriched se-quences are also detected in the contralateral breast.Pearson correlation coefficients of the abundance from the

top 100 clonotypes between each pair of tissues (SI Appendix,Figs. S10 and S11) indicate that normal breast, blood, and con-tralateral breast are more correlated with each other (0.45–0.6)than with the tumor (−0.13 to −0.03; SI Appendix, Fig. S8A).This trend is consistent across many of the patients, with larger

A B C

D E F

Fig. 1. T cell abundance and repertoire diversity in blood, tumor, and adjacent normal breast tissue. (A) Matched trio of samples was obtained from bloodPBMCs, tumor, and normal breast tissue for each patient. Genomic DNA was isolated from each sample, and the TCRB region was PCR-amplified beforeIllumina sequencing. The set of unique clonotypes, defined as TCRB sequences with the same V family, J gene, and CDR3 amino acid sequence, and thenumber of input template molecules for each clonotype comprise the TCRB repertoire in each sample. The number of T cells in each tissue is estimated fromthe abundance of in-frame TCRB template molecules (B), and the number density is determined as the fraction of templates relative to the number of inputcells (C) (SI Appendix, Table S1). Repertoire diversity for each tissue is estimated from the fraction of templates corresponding to unique clonotypes (D), thecumulative abundance of the top 50 clonotypes in each sample (E), and the clonality (F). Additional details are provided in Materials and Methods. Unlessotherwise indicated, *P < 0.05; **P < 0.005; ***P < 0.0005. B, blood; N, normal breast; T, tumor.

Beausang et al. PNAS | Published online November 14, 2017 | E10411

IMMUNOLO

GYAND

INFLAMMATION

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

correlations between normal breast and blood (median correla-tion of 0.40) compared with normal breast and tumor (mediancorrelation of 0.08) or blood and tumor (median correlation of0.01; SI Appendix, Fig. S8B).

Classifying Tumor and Normal Breast from the T Cell Repertoire. Thedifferences in clonal structure between tumor and normal breast(Fig. 1) suggest that the T cell repertoire could be used toidentify a hypothetical biopsy of uncertain tissue as either normalbreast or tumor. The larger correlation between normal breastand blood compared with tumor and blood (SI Appendix, Fig.S8B) indicates that including a matched blood sample along witha tissue biopsy would further improve the performance of suchan assay. We find that the fraction of the variance in the datacaptured by a linear model fit to the abundance of the top100 clonotypes in blood with the corresponding abundance ineither tissue [i.e., the coefficient of determination (R2) of the fitlines in SI Appendix, Figs. S10 and S11] distinguishes tumor(median of 0.17) from normal breast (median of 0.49; P < 3.5e5;Fig. 4A). Receiver operator curves comparing performance ofthe R2 metric with T cell density, clonality, and unique T cellfraction show that all four have reasonable performance but thatthe R2 metric has a peak sensitivity and specificity of ∼0.94 and alarger area under the curve (AUC) of 0.98 compared with theothers (AUCs of 0.85, 0.74, and 0.79, respectively; Fig. 4B).

Subsets of Sequences Are Highly Enriched in Tumors and in NormalBreast Tissues. The number of tumor-enriched sequences variesacross patients from five to 64 (median of 22) sequences persample, corresponding to 0.9–29.4% (median of 6.7%) of tem-plates in each tumor (Fig. 5A). While the exact value of thethreshold used to define enriched is somewhat arbitrary, the dis-tribution of relative abundances indicates that the proportion ofclonotypes with tumor/normal breast greater than 32 exceeds theamount expected from a log-normal distribution fit to the binneddata (Fig. 5C). There also exists a similar distribution of enriched

sequences in normal breast tissue where the number of normalbreast-enriched sequences per sample varies between 1 and 65(median of 18), with template abundances ranging from 0.9 to29% (median of 6.7%; Fig. 5B). The overall number of normalbreast-enriched sequences also exceeds what is expected from alog-normal fit (Fig. 3D). In contrast, the number of enriched se-quences with abundance less than 0.1% in tumor and normalbreast is relatively low and does not exceed the amount predictedfrom a log-normal fit to the distribution (SI Appendix, Fig. S12).Across all patients, ∼40% of tumor-enriched and 32% of nor-

mal breast-enriched clonotypes were detected in peripheral blood.For tumor-enriched sequences, there is no correlation betweenthe abundances in tumor and blood, with only two of 370 tumor-enriched sequences detected above 0.05% abundance in blood(Fig. 3E). In normal breast, however, 44 of 761 (6%) enrichedsequences exceed 0.05% abundance in blood and are highly cor-related with the abundance in normal tissue (Fig. 3F). Thus,normal breast-enriched sequences seem to have two subsets, onewith abundances that are correlated with those in blood and theother where they are not. Tumor-enriched clonotypes only containthe subset of sequences that are uncorrelated with the abundancein blood.

Shared Sequences Between Patients Are Consistent with CommonSequences. The subset of TCRB CDR3 sequences sharedacross multiple samples may contain clonotypes that recognize acommon shared antigen, such as a breast cancer epitope, or se-quences that occur naturally across multiple people. We showbelow that commonly shared CDR3 sequences have similar

A B

C D

Fig. 3. Scatter plots of clonotype abundance for patient BR21. (A) Tumor(T) vs. normal breast (N). (B) Tumor vs. blood. (C) Normal breast vs. blood.(D) Normal breast vs. contralateral normal breast. Each point represents aclonotype detected in one (diamonds) or both (squares) samples. Clonotypeswith equal abundance in both samples lie along the diagonal (solid blackline), whereas clones enriched in the tumor relative to normal breast ornormal relative to tumor are defined as relative abundance greater than32 and absolute abundance greater than 0.1% (dashed lines). Note thatmany sequences, especially at low abundance, contain the same number oftemplates in both samples but are only represented by a single point (thedistribution of clonotype abundances is shown in SI Appendix, Fig. S7).

B T N

B - 0.52 0.50

T 0.30 - 0.54

N 0.26 0.47 -

B T N

B 0.0071 0.0080 0.0073

T 0.0028 0.0041 0.0036

N 0.0022 0.0031 0.0028

ji

ji

A B

C D

Fig. 2. Overlap of TCRB templates between tissues and patients. Theoverlap fraction between two samples is defined as the fraction of templatesin tissue j (columns) with sequences that are also detected in tissue i (rows),with the median value reported for within (A) and between (B) patienttissue combinations. (C) On average (thick gray line) and individually (thingray lines), sequences with high abundance in the tumor are more likely tobe detected in normal breast. Similar trends hold for other tissue combina-tions (SI Appendix, Fig. S4). (D) Overlap between samples observed for eachpatient is strongly correlated (linear regression with 95% confidence intervaland Pearson correlation coefficient, r) with the clonality of the patient’srepertoire. Avg., Average; B, blood; N, normal breast; T, tumor.

E10412 | www.pnas.org/cgi/doi/10.1073/pnas.1713863114 Beausang et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

properties in the three tissue compartments and, unlike tumor-enriched sequences, are consistent with commonly occurringCDR3 sequences in a population of healthy donors.The fraction of clonotypes with shared CDR3 sequences be-

tween two patients across the different tissue combinations iswell described (within ∼20%) by a standard capture-recapturemodel with a single-fit parameter representing the total numberof productive sequences in the population (M = 2 × 106; SIAppendix, Fig. S13). To compare the degree of sharing betweendifferent tissues, we compare the overlap among the top∼2,000 most abundant CDR3 sequences in each tissue (Materialsand Methods). Using the inferred population size M, we simu-lated the expected fraction of sequences detected in multiplepatients and show that the amount of sharing exceeds what isexpected from random sampling (Fig. 6A). Approximately 2% ofabundant tumor and normal breast sequences and 4% of those inblood are shared across two or more patients. In particular, thetail of the distribution is larger than expected, with 102, 13, and6 highly shared sequences detected in more than five patientsfrom blood, tumor, and normal breast, respectively. The capture-recapture model ignores clonotype abundance within each pa-tient, but this assumption is consistent with the minimal biasobserved in the abundance of highly shared CDR3 sequencescompared with unshared sequences (Fig. 6B).Several metrics indicate that highly shared sequences are less

diverse than private sequences. First, average CDR3 length isshorter for shared sequences, decreasing from 14.5 aa in non-shared sequences to ∼11 aa in the most highly shared sequences(Fig. 6C). The number of nucleic acid bases deleted from theCDR3 region of shared sequences is relatively unchanged (Fig.6D), but the number of inserted bases in the CDR3 region de-creases sharply from approximately six nucleic acids in non-shared sequences to approximately one nucleic acid in sharedsequences (Fig. 6E). Lastly, the edit distance between all pairs ofrandomly selected nonshared sequences with a length of 13 aahas a relatively smooth distribution that is peaked at eight aminoacid bases. In contrast, the edit distance distribution betweenshared CDR3 sequences is shifted toward shorter lengths, with ashoulder at approximately three amino acid bases indicating aless diverse subset (Fig. 6F).We used a simple model to compute low-diversity recombina-

tions from all germline amino acid contributions from the V-, D-,and J-gene segments to the CDR3 region. The artificial CDR3sinclude amino acid deletions, but no insertions, and result in116,367 unique sequences (Fig. 6G, further details provided inMaterials and Methods). Eighty percent of these are not detected

in our study, mostly due to nonuniform V- and J-gene usage inactual repertoires; 10% were detected in nonshared clonotypes,with the remaining 10% distributed across all shared CDR3s.After subsampling, the fraction of shared CDR3s predicted by themodel increases rapidly with the degree of sharing, with 100% ofthe most frequently observed clonotypes in each tissue predictedby the model (Fig. 6H).Lastly, shared CDR3 sequences in tumor, blood, and normal

breast are almost all relatively common in a database of T cellrepertoires from 585 healthy donors (18) (Fig. 6I). In contrast,tumor-enriched sequences are much less frequently observed inthe database, with ∼50% of CDR3s detected in less than fivedonors. Both tumor-enriched and shared CDR3 sequences aresimilarly represented in male and female donors (SI Appendix,Fig. S15), which is expected for the commonly shared sequencesbut not for enriched sequences, suggesting that the most highlyenriched T cells in the tumor may not be recognizing antigensspecific to female breast cancer.

A B

C D

E F

Fig. 5. Enriched sequences in tumor and normal breast. The total numberof sequences with abundance greater than 0.1% in each patient, includingthe number that are enriched in the tumor relative to normal breast (A) andin the normal breast relative to tumor (B), is shown The relatively largenumber of enriched sequences in normal breast from BR18 is likely due tothe low-input amount of normal breast DNA in this sample (SI Appendix,Table S1), which can be seen in the small number of low-abundance (<0.1%)sequences (SI Appendix, Fig. S12). The distribution of enriched sequencesacross all patients for sequences in the tumor (C) and normal breast (D, notincluding the outlier BR18), including a curve fit to a log normal distribution(solid black line), is shown. N, normal breast; T, tumor. Scatter plots comparethe enriched clonotypes in tumor (E) and normal breast (F) with theirabundance in peripheral blood. Both high- and low-abundance (0.1%, ver-tical dashed line) clonotypes are shown, and the fractions of sequences thatare highly correlated with blood (solid diagonal line) and with abundancegreater than 0.05% (horizontal dashed line) are enumerated. Clonotypesdetected in tumor or normal breast but not detected (n.d.) in blood (♢) aredepicted with arbitrary tissue abundance.

A B

Fig. 4. Distinguishing between TCRB repertoires from tumor and normalbreast. (A) R2 resulting from fitting the tumor vs. blood abundance andnormal breast vs. blood abundance (fit lines are shown in SI Appendix, Figs.S10 and S11). ***P < 0.0005. (B) Receiver operator curves and correspondingAUC values for metrics that distinguish repertoires between tumor andnormal breast. B, blood; Frac., fraction; N, normal breast; T, tumor.

Beausang et al. PNAS | Published online November 14, 2017 | E10413

IMMUNOLO

GYAND

INFLAMMATION

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

DiscussionIn this work, we performed bulk T cell repertoire sequencingfrom peripheral blood, adjacent normal breast, and tumor from16 prospectively collected patients with breast cancer, most ofwhich were early-stage, treatment-naive, ER+/PR+/HER2− sub-type tumors. Advantages of bulk repertoire sequencing includeits high sensitivity for detecting T cell clonotypes (16), which maybe missed by traditional staining (29), and the simplicity of ge-nomic DNA input, which is not subjected to the variability ofsingle-cell isolation techniques (33) required in cytometry. Thedisadvantage is that some important cellular details are missed,such as the heavy- and light-chain pairing and the expression offunctional markers, such as CD4, CD8, FOXP3, CTLA4, andPDL1 (19, 34), unless cells have been previously sorted (15, 35).

Our goal was to characterize the differences in repertoires be-tween adjacent normal breast tissue and tumor to determine ifthere was evidence for subsets of T cells that might play a role inthe tumor microenvironment (36).Recent guidelines have been established for evaluating TILs in

breast cancer via standard hematoxylin and eosin (H&E) staining(37), and clinical studies have shown that outcomes are correlatedwith the number of infiltrating T cells, especially for triple-negative and HER2+ subtypes (38). Quantifying T cells via se-quencing is highly reproducible and correlates well with H&Estaining but with higher sensitivity (29). Immunohistochemistryindicates that T cells in healthy breast tissue are mostly located inlobules, with CD8+ T cells ∼10-fold more abundant than CD4+

T cells (39). Immune infiltrates isolated from tumor and normal

A B C

D E F

G H I

Fig. 6. Interpatient T cell sharing of clonotypes between tissues. (A) Fraction of blood, tumor, and normal breast tissue sequences found in one (not shared)or more (shared) patients compared with the expected fraction by random sampling (dashed lines). (B) Median abundance of clonotypes detected in multiplepatients is not sensitive to the degree of sharing. Compared with clonotypes that are not shared across multiple patients, shared clonotypes have a shorterCDR3 amino acid (AA) length (C), which can be attributed to a similar number of deleted (D) but fewer inserted (E), nucleotide bases in the junctional regionsbetween V-D and D-J genes. Avg., Average. (F) Diversity of CDR3 sequences is estimated from the distribution of edit distances (length = 13 aa), which isshifted to lower values for shared sequences. Number and length of germline V-, D-, and J-gene amino acid sequence fragments used in a recombinationmodel of low-diversity CDR3 sequences (G) and the fraction of CDR3 sequences detected by the model (H) are shown (details are provided in Materials andMethods). (I) Results from a database query of 585 healthy donors (18) indicate that shared CDR3 sequences are also likely to be shared in a population ofhealthy donors, whereas enriched CDR3 sequences are much less likely to be shared in the general population.

E10414 | www.pnas.org/cgi/doi/10.1073/pnas.1713863114 Beausang et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

breast contain similar amounts of CD8+ T cells (∼20% of T cells),whereas CD4+ T cells in tumor comprise ∼40% of T cells com-pared with ∼20% in normal breast tissue (40). We show thatrepertoires from PBMCs, tumor, and normal breast are distinct,with large differences in the number and density of T cells de-tected in each tissue compartment. The tumor repertoire is lessdiverse than in normal breast, as indicated by the lower fraction ofunique T cells and increased clonality in the tumor (Fig. 1 E andF). These results are consistent with a similar study comparing theeffects of cryoablation and immunotherapy on TCRB repertoiresin patients with early-stage breast cancer (29).These repertoire metrics have predictive value and could be

used to classify unknown tissue samples taken from tumor ornormal breast. The performance in this small cohort varies froman AUC of 0.74 for clonality to an AUC of 0.85 for T cell density.Since repertoires from normal breast and blood are more similarthan those from tumor and blood, a better metric can be de-termined by comparing the tissue and blood repertoires. Usingthe R2 goodness of fit between each tissue with blood, the AUCincreases to 0.98 with only two samples not properly classified.Extensive overlap between tissues within the same patient is

expected due to normal blood perfusion. On average ∼50% ofT cells in either tissue sample were found to overlap with theother tissue or blood samples from the same patient. This frac-tion varied from patient to patient and was correlated with theaverage clonality of the repertoires in each patient, which isconsistent with the more clonal repertoires containing largerclonotypes that are more frequently sampled in different tissues.In addition to many frequently overlapping clonotypes, most

patient tumors contained a subset of clonotypes (median of 22)that are highly enriched in the tumor relative to normal breast.In a few patients (BR05, BR21, and BR22 in SI Appendix, Fig.S9), these clonotypes are clearly dominant in the tumor; how-ever, in most patients, high-abundance clonotypes in the tumorare also highly abundant in the normal breast, suggesting thatthey are less likely to play an important role in the tumor mi-croenvironment. Methods have been developed to identify thesubset of clonotypes that change abundance in serial bloodsamples following an external perturbation (41). Analogoussubsets are not as clear between breast tumors and normal tissue,but there is a population of highly enriched clonotypes in eachtissue type across the cohort of patients (Fig. 5 C and D). Wefocused our analysis of enriched clonotypes on those with at least0.1% abundance due to limited sequencing depth and multiplereports that tumor-reactive T cells identified in other cancerstypically exceed 0.1%. For example, in melanoma and early-stageNCSLC, ∼3% (42) and 0.2–1.5% (43) of infiltrating CD8+

T cells are tumor-reactive, with sorting on PD1+ in melanomafurther enriching the population to 1–10% (44). CD4+ T cellsmay also be tumor-reactive (45, 46), with 0.13–0.3% of expandedCD4+ T cells being tumor-reactive in melanoma (47).Enriched clonotypes in normal breast tissue show two subsets in

peripheral blood: one that is highly correlated with blood and theother distributed over a range of low (or zero) abundance inblood. We interpret the first subset as sequences from blood thatperfuse normal tissue, whereas the second subset may represent acompartment of tissue-resident T cells in normal breast (48). Theabundance distribution of tumor-enriched clonotypes in bloodresembles this second subset seen in normal breast and does notcontain clonotypes that are highly correlated with blood. Whilesome of the tumor-enriched T cells may play a direct role in thetumor microenvironment, many T cells in this population mayinstead represent a compartment of tissue-specific resident T cellsthat are distinct from blood and performing standard immunesurveillance (49).Public T cells expressing the same TCR sequence but gener-

ated in different individuals have been identified for numerousinfectious diseases, autoimmune disorders, and some cancers

(50). In cancer, this process is thought to require T cells thatrecognize epitopes from mutated proteins that are presented bysimilar major MHC molecules (3), and thus occur more often incancers with larger numbers of mutations (8). We detect a smallfraction of TCRB CDR3s across multiple tumors, but ouranalysis suggests most of these are low-diversity sequences thatare unlikely to be tumor-specific. First, CDR3 sequences sharedacross tumors follow similar trends as those shared in blood andnormal breast, with a similar amount detected in normal breastas in tumor. Shared clonotypes were not strongly biased byabundance and contained shorter CDR3 lengths with manyfewer inserted nucleotide bases into the junction regions, asreported in earlier studies (15, 51) and recently reviewed (52). Asimple model generating all CDR3 recombinations from germ-line amino acid sequences with minimal insertions was able topredict many of the shared sequences, including all of the mosthighly shared sequences. These findings are consistent with moredetailed physical models of the nucleotide sequence recombi-nation (53) that show commonly occurring sequences have ahigher likelihood of being generated.Shared and tumor-enriched sequences are distinct, with only

two CDR3s detected in both sets. We also queried a large publicdatabase of T cell repertoires in blood from 585 healthy volunteers(18). Enriched sequences were detected in many fewer healthydonors than shared sequences, supporting the hypothesis thatenriched sequences may play a specialized role in a subset ofpatients, whereas most shared sequences are common in thepopulation. Interestingly, tumor-enriched sequences were equallyprevalent in male and female volunteers, suggesting that they maynot be specific to breast cancer.

ConclusionsBy comparing bulk T cell repertoires from tumor and normalbreast tissue in a cohort of patients with early-stage breast can-cer, we were able to determine the clonal structure of infiltratinglymphocytes. These clonal structures were quite distinct betweentumor and normal breast, and we were able to distinguish tissuesas normal or tumor solely on the basis of comparing T cellrepertoire with blood. We identified a subset of T cells in eachpatient that were highly abundant in the tumor compared withnormal breast; this population may represent a clinically relevantsubset for future investigation. In many patients, however, theseclonotypes were dominated by more abundant clonotypes, whichwere also highly abundant in normal breast, thus highlighting thelarge amount of background T cells in the tumor that complicateisolating and identifying tumor-specific T cells.

Materials and MethodsSample Collection. Informed consent was obtained from patients with in-vasive breast cancers greater than 1 cm under Stanford Institutional ReviewBoard-approved Research Protocol 5630. Whole blood (5–10 mL) was col-lected in EDTA tubes before surgery and processed within 2 h. After tumorresection, tumor tissue was visually identified and a portion was excised andplaced in prechilled RNAlater. Samples of adjacent normal breast tissue wereobtained from sites 2–4 cm away from the tumor, rinsed with PBS, andplaced in a separate tube of prechilled RNAlater. For one patient un-dergoing simultaneous double mastectomy, normal breast tissue was alsoobtained from the contralateral breast.

Sample Processing. Plasma was removed after separation by centrifugationfor 10 min at 1,600 × g. The remaining blood cells were resuspended in PBS,and mononuclear cells were isolated via Ficoll-Paque centrifugation withSepmate-50 tubes (Stem Cell Technologies). Cells were cryopreserved in 90%FBS and 10% DMSO, divided into aliquots containing 2 to 4 million cells, andstored in liquid nitrogen until use. Cells were thawed at room temperature,diluted with PBS, pelleted at 350 RCF, and resuspended in RLTplus buffer(Qiagen) with 1% beta-mercaptoethanol (Sigma) and 0.05% Reagent DX anti-foaming agent (no. 19088; Qiagen). Lysate was passed through a QIAshredder(Qiagen), and DNA and RNA were extracted using an AllPrep DNA/RNA Minikit (Qiagen) following the manufacturer’s instructions.

Beausang et al. PNAS | Published online November 14, 2017 | E10415

IMMUNOLO

GYAND

INFLAMMATION

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

DNA Isolation. Tumor and normal breast tissue samples were divided intopieces, incubated in RNAlater overnight, and transferred to −20 °C the followingday. For the tumor, 20–50 mg of tissue was diced with a fresh razor blade andhomogenized in RLTplus buffer containing 1% beta-mercaptoethanol and 0.5%reagent DX antifoaming agent using a Tissuelyzer II (Qiagen) with oscillationfrequency set to 30/s for 3–9 min. Homogenized lysate was centrifuged at20,000 RCF for 3 min, and the supernatant was removed and passed througha QIAshredder before proceeding with the Allprep Mini kit for DNA and RNAisolation. An on-column digestion with 20 μL of proteinase-K (Qiagen) dis-solved in 60 μL of buffer AW1 was implemented before washing the DNAcolumn with buffer AW2. DNA yield from tumors was variable, but, typically,20–50 mg of tumor was required for >3 μg of DNA. Normal breast tissue wasprocessed similarly except 100–300 mg of tissue was often required to obtainsufficient (>3 μg) DNA due to high lipid content and low density of cells. Inparticular, care was taken to avoid the lipid layer after the 20,000 × g cen-trifugation step. DNA was quantified with fluorometric quantitation (QubitHigh Sensitivity DNA kit) and UV spectroscopy (Nanodrop 1000) (SI Appen-dix, Table S2). To minimize any contamination, samples were processed atseparate times and all work was performed in a PCR cabinet (AirClean).

Immunosequencing. TCRB sequencing was performed on genomic DNA pu-rified from blood, tumor, and normal breast tissue using the ImmunoSeqIntro kit (Adaptive Biotechnologies) following the manufacturer’s instruc-tions. Briefly, two genomic DNA replicates per sample, each containing ei-ther 0.5 μg (blood) or 1.5 μg (tumor, normal breast; details are provided in SIAppendix, Table S1) are independently amplified using 2× Multiplex PCRMaster mix (31 cycles; no. 206151; Qiagen) with proprietary primers andspike-in control sequences for absolute quantification (13, 54). The productis cleaned up with AMPure XP magnetic beads, and amplified again witheight additional cycles of PCR to attach Illumina sequencing adapters con-taining sample-specific barcodes. An 87-bp fragment that includes theCDR3 region and flanking V and J genes is sequenced using single-ended150-bp reads of seven samples (including two replicates for each sample)multiplexed onto a single lane of an Illumina MiSeq sequencer with version3.0 chemistry. The raw sequencing data are uploaded to Adaptive Biotech-nologies and processed using the company’s proprietary pipeline. Results foreach sample are reported to the user after merging data from both repli-cates and include absolute quantification of template molecules for allnucleic acid sequences detected; CDR3 amino acid sequence for in-framemolecules; V-, D-, and J-gene segment identification; the most likely num-ber of bases deleted from each gene segment contributing to the CDR3; andthe most likely number of bases inserted into either junction. The averageread coverage of each sample was ∼10-fold (details are provided in SI Ap-pendix, Table S2).

Data Analysis. Processed sequence data for each sample were downloadedfrom Adaptive Biotechnologies and analyzed with custom scripts in bash,awk, and R. Unless otherwise noted, clonotypes are defined here as pro-ductive recombinations containing a V-gene family, CDR3 amino acid se-quence, and J-gene segment. Sequence abundance is calculated as thenumber of templates for that sequence divided by the total number oftemplates in the sample. After normalization via the spike-in control se-quence, the assay is reported to be quantitative to one T cell in 20,000 (54).The T cell density is determined by dividing the total number of templatesby the total number of input cells estimated from the input amount ofDNA. Clonality represents the distribution of clone sizes and is defined as1 minus the normalized Shannon entropy of the TCRB abundances (i.e., thePielou evenness; SI Appendix, Table S1). Using this definition, clonalityranges from zero, where all T cells are evenly distributed across all clono-types, to unity, where all T cells are represented by a single clonotype. Pvalues between tissues are calculated by a paired Wilcoxon rank sum test inR unless otherwise noted, with *, **, and *** indicating P values less than0.05, 0.005, and 0.0005, respectively. Estimates of diversity in unique clo-notypes and clonality were performed using Recon analysis with defaultparameters (32).

Overlapping Templates. The fraction fk’ ,k

i,j of templates in patient k, tissuej that are shared with patient k′, tissue i is defined as the sum of theabundances of all sequences in sample ( j,k) that were also detected at least

once in (i,k′). As a result, fk’ ,k

i,j ≠ fk,k’

j,i and the fraction of overlapping sequences

between tissues with large differences in total numbers of sequences can berepresented (SI Appendix, Fig. S3). Intrapatient and interpatient averages aredetermined by finding the median over tissue pairs (i,j) when k = k′ and k ≠ k′,respectively (Fig. 2 A and B).

Capture-Recapture Model. The number of overlapping sequences ~nk’ ,ki,j be-

tween tissues (i,j) is modeled as

~nk’ ,ki,j =

Nk’i N

kj

M  ði   and  j   between  patients,  k  ≠   k’

�,

where Nk’i and Nk

j are the unique sequences observed in each tissue, and Mrepresents the total number of unique sequences in the study populationand is determined by minimizing the ratio, g, of modeled and measuredcounts over all pairs of tissues between patients (k ≠ k′):

gi,j =

*~nk’ ,ki,j

nk,k’i,j

+k≠k’

.

Results and a detailed comparison of data to the model are provided in SIAppendix, Fig. S13.

Tumor-Normal Classifier. Receiver operator curves and the AUC were com-puted with the ROCR package in R using the density of T cells, fraction ofunique clonotypes, and clonality (Fig. 1) as metrics to distinguish repertoiresfrom tumor and normal tissue. A classifier based on the R2 value comparingthe top 100 clones in blood with tumor and normal breast was also evalu-ated. Linear regression was performed between the abundances of the top100 clonotypes in blood and the corresponding abundances in tumor, andthe fraction of variance captured by the fit (R2) was tabulated for eachsample (SI Appendix, Fig. S10). The same process was repeated between thetop 100 clonotypes in blood with normal breast (SI Appendix, Fig. S11).

Enriched Sequences. Sequences between pairs of tissues are plotted on log–log plots, where sequences with zero counts in one of the samples areplotted at 1/4 of the lowest abundance in that sample. Tumor sequence n isenriched if it has abundance greater than 0.1% and relative abundance inthe normal breast exceeding 32 (i.e., Tn/Nn ≥ 32), including clonotypes notdetected in normal breast (i.e., Nn = 0). Similarly, enriched normal sequencesare defined as sequences with greater than 0.1% abundance in the normalbreast and a ratio with the tumor exceeding 32 (i.e., Nn/Tn ≥ 32).

Interpatient Sequence Sharing. The number of patients with at least onetemplate detected for each CDR3 sequence is independently tallied for thethree tissue compartments, and the sequences are binned according to thenumber of shared patients (SI Appendix, Fig. S14). To more easily comparesharing between samples and tissues, this process is repeated for the top 1,945CDR3s from each patient (only BR18N has fewer sequences). The expectedamount of sharing (dashed lines in Fig. 6A and SI Appendix, Fig. S14) is de-termined by random sampling sequences from a population of size M = 2 ×106 according to the sampling depth in each patient and tallying the numberof common clonotypes across patients. The median abundance, CDR3 length,and number of inserted/deleted nucleotide bases in the junction are computedfor each bin. For sequences where the D gene could not be resolved, thenumber of deleted bases is conservatively estimated at 10 deleted nucleotides,which is the maximum number of deleted D-gene bases reported. The editdistance between two CDR3 sequences was calculated as the number of aminoacid base changes required to transform one CDR3 sequence into the other.The distribution of edit distances between all pairs of CDR3 sequences with alength of 13 aa from shared sequences (detected in more than one patient)was compared with nonshared sequences (only detected in one patient).

CDR3 Recombination Model. Commonly shared low-diversity sequences aregenerated from the international ImMunoGeneTics information system(IMGT) germline amino acid sequences for the human V, D, and J genes.Amino acid sequences contributing to the CDR3 region are extracted fromthe 5′-end of each V gene starting with the conserved cysteine (or nearestequivalent in some V genes), the 3′-end of each J gene ending in the con-served phenylalanine, and the full D gene, including all three readingframes. All combinations of these sequences are then used to generate a listof low-diversity CDR3 sequences. To reflect our observation that sharedCDR3s are shorter with many deletions and few insertions, all possible aminoacid truncations of the D gene and one to two amino acid truncations fromthe 3′-end of the J gene are also included. No inserted amino acids are in-cluded. In general, nucleic acid sequences are ignored except when eitherend of the D gene or the 3′-end of the J gene contains two nucleic acids, thusstrongly biasing the amino acid usage at this position. These additionalamino acids are not strictly germline, but are also included. A detailed list ofsequences from each gene is provided in SI Appendix, Table S3.

E10416 | www.pnas.org/cgi/doi/10.1073/pnas.1713863114 Beausang et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020

Database of Healthy Donors. To query the database of 585 healthy bonemarrow donors (18), a list of the top 400 shared and 858 enriched CDR3sequences was compiled. Shared sequences include those CDR3s that aredetected in at least two tumor samples, two normal breast samples, or fourblood samples. These sequences were provided to Adaptive Biotechnologies,where a query on their internal database was performed, and the subset ofImmunoSeq data from each patient’s matching clonotypes was provided tous. For each TCRB sequence, the number of patients, the relative abun-dances, and the corresponding nucleic acid sequences were tallied forenriched/shared sequences and male/female donors.

Data Availability. All immunosequencing data underlying this study can be an-alyzed and freely downloaded from theAdaptiveBiotechnologies immuneACCESSsite at https://clients.adaptivebiotech.com/pub/beausang-2017-pnas.

ACKNOWLEDGMENTS. We thank Yasemin Sucu for help in processing sam-ples, Dr. David Hamm and his team at Adaptive Biotechnologies for assis-tance in the database query, and Florian Rubelt for helpful discussions. Thisstudy received funding from the Howard Hughes Medical Institute, MattiasWestman, the John and Marva Warnock Research Fund, and the Debra andAndrew Rachleff Research Fund.

1. Dunn GP, Bruce AT, Ikeda H, Old LJ, Schreiber RD (2002) Cancer immunoediting: Fromimmunosurveillance to tumor escape. Nat Immunol 3:991–998.

2. Fridman WH, Pagès F, Sautès-Fridman C, Galon J (2012) The immune contexture inhuman tumours: Impact on clinical outcome. Nat Rev Cancer 12:298–306.

3. Heemskerk B, Kvistborg P, Schumacher TNM (2013) The cancer antigenome. EMBO J32:194–203.

4. Blankenstein T, Coulie PG, Gilboa E, Jaffee EM (2012) The determinants of tumourimmunogenicity. Nat Rev Cancer 12:307–313.

5. Chen DS, Mellman I (2013) Oncology meets immunology: The cancer-immunity cycle.Immunity 39:1–10.

6. Pardoll DM (2012) The blockade of immune checkpoints in cancer immunotherapy.Nat Rev Cancer 12:252–264.

7. Coulie PG, Van den Eynde BJ, van der Bruggen P, Boon T (2014) Tumour antigensrecognized by T lymphocytes: At the core of cancer immunotherapy. Nat Rev Cancer14:135–146.

8. Schumacher TN, Schreiber RD (2015) Neoantigens in cancer immunotherapy. Science348:69–74.

9. Dushyanthen S, et al. (2015) Relevance of tumor-infiltrating lymphocytes in breastcancer. BMC Med 13:202.

10. Kwa M, Adams S (2016) Prognostic and predictive value of tumor-infiltrating lym-phocytes in breast cancer. Curr Breast Cancer Rep 8:1–13.

11. Miyan M, Schmidt-Mende J, Kiessling R, Poschke I, de Boniface J (2016) Differentialtumor infiltration by T-cells characterizes intrinsic molecular subtypes in breast can-cer. J Transl Med 14:227.

12. Savas P, et al. (2016) Clinical relevance of host immunity in breast cancer: From TILs tothe clinic. Nat Rev Clin Oncol 13:228–241.

13. Robins HS, et al. (2009) Comprehensive assessment of T-cell receptor β-chain diversityin alphabeta T cells. Blood 114:4099–4107.

14. Freeman JD, Warren RL, Webb JR, Nelson BH, Holt RA (2009) Profiling the T-cell re-ceptor beta-chain repertoire by massively parallel sequencing. Genome Res 19:1817–1824.

15. Robins HS, et al. (2010) Overlap and effective size of the human CD8+ T cell receptorrepertoire. Sci Transl Med 2:47ra64.

16. Robins H, et al. (2012) Ultra-sensitive detection of rare T cell clones. J ImmunolMethods 375:14–19.

17. Wu D, et al. (2012) High-throughput sequencing detects minimal residual disease inacute T lymphoblastic leukemia. Sci Transl Med 4:134ra63.

18. Emerson RO, et al. (2017) Immunosequencing identifies signatures of cytomegalovirusexposure history and HLA-mediated effects on the T cell repertoire. Nat Genet 49:659–665.

19. Han A, Glanville J, Hansmann L, Davis MM (2014) Linking T-cell receptor sequence tofunctional phenotype at the single-cell level. Nat Biotechnol 32:684–692.

20. Munson DJ, et al. (2016) Identification of shared TCR sequences from T cells in humanbreast cancer using emulsion RT-PCR. Proc Natl Acad Sci USA 113:8272–8277.

21. Howie B, et al. (2015) High-throughput pairing of T cell receptor α and β sequences.Sci Transl Med 7:301ra131.

22. Kirsch I, Vignali M, Robins H (2015) T-cell receptor profiling in cancer. Mol Oncol 9:2063–2070.

23. Sherwood AM, et al. (2013) Tumor-infiltrating lymphocytes in colorectal tumors dis-play a diversity of T cell receptor sequences that differ from the T cells in adjacentmucosal tissue. Cancer Immunol Immunother 62:1453–1461.

24. Gerlinger M, et al. (2013) Ultra-deep T cell receptor sequencing reveals the complexityand intratumour heterogeneity of T cell clones in renal cell carcinomas. J Pathol 231:424–432.

25. Chen Z, et al. (2016) T cell receptor β-chain repertoire analysis reveals intratumourheterogeneity of tumour-infiltrating lymphocytes in oesophageal squamous cellcarcinoma. J Pathol 239:450–458.

26. Emerson RO, et al. (2013) High-throughput sequencing of T-cell receptors reveals ahomogeneous repertoire of tumour-infiltrating lymphocytes in ovarian cancer.J Pathol 231:433–440.

27. Bai X, et al. (2015) Characteristics of tumor infiltrating lymphocyte and circulatinglymphocyte repertoires in pancreatic cancer by the sequencing of T cell receptors. SciRep 5:13664.

28. Zhu W, et al. (2015) A high density of tertiary lymphoid structure B cells in lung tu-mors is associated with increased CD4+ T cell receptor repertoire clonality.Oncoimmunology 4:e1051922.

29. Page DB, et al. (2016) Deep sequencing of T-cell receptor DNA as a biomarker ofclonally expanded TILs in breast cancer after immunotherapy. Cancer Immunol Res 4:835–844.

30. Li B, et al. (2016) Landscape of tumor-infiltrating T cell repertoire of human cancers.Nat Genet 48:725–732.

31. Levy E, et al. (2016) Immune DNA signature of T-cell infiltration in breast tumorexomes. Sci Rep 6:30064.

32. Kaplinsky J, Arnaout R (2016) Robust estimates of overall immune-repertoire diversityfrom high-throughput measurements on samples. Nat Commun 7:11881.

33. Steinert EM, et al. (2015) Quantifying memory CD8 T cells reveals regionalization ofimmunosurveillance. Cell 161:737–749.

34. Chung W, et al. (2017) Single-cell RNA-seq enables comprehensive tumour and im-mune cell profiling in primary breast cancer. Nat Commun 8:15081.

35. Qi Q, et al. (2014) Diversity and clonal selection in the human T-cell repertoire. ProcNatl Acad Sci USA 111:13139–13144.

36. Gajewski TF, Schreiber H, Fu Y-X (2013) Innate and adaptive immune cells in the tu-mor microenvironment. Nat Immunol 14:1014–1022.

37. Salgado R, et al.; International TILs Working Group 2014 (2015) The evaluation oftumor-infiltrating lymphocytes (TILs) in breast cancer: Recommendations by an In-ternational TILs Working Group 2014. Ann Oncol 26:259–271.

38. Denkert C, et al. (2015) Tumor-infiltrating lymphocytes and response to neoadjuvantchemotherapy with or without carboplatin in human epidermal growth factor re-ceptor 2-positive and triple-negative primary breast cancers. J Clin Oncol 33:983–991.

39. Degnim AC, et al. (2014) Immune cell quantitation in normal breast tissue lobuleswith and without lobulitis. Breast Cancer Res Treat 144:539–549.

40. Ruffell B, et al. (2012) Leukocyte composition of human breast cancer. Proc Natl AcadSci USA 109:2796–2801.

41. DeWitt WS, et al. (2015) Dynamics of the cytotoxic T cell response to a model of acuteviral infection. J Virol 89:4517–4526.

42. van Rooij N, et al. (2013) Tumor exome analysis reveals neoantigen-specific T-cellreactivity in an ipilimumab-responsive melanoma. J Clin Oncol 31:e439–e442.

43. McGranahan N, et al. (2016) Clonal neoantigens elicit T cell immunoreactivity andsensitivity to immune checkpoint blockade. Science 351:1463–1469.

44. Pasetto A, et al. (2016) Tumor- and neoantigen-reactive T-cell receptors can beidentified based on their frequency in fresh tumor. Cancer Immunol Res 4:734–743.

45. Kreiter S, et al. (2015) Mutant MHC class II epitopes drive therapeutic immune re-sponses to cancer. Nature 520:692–696.

46. Tran E, et al. (2014) Cancer immunotherapy based on mutation-specific CD4+ T cells ina patient with epithelial cancer. Science 344:641–645.

47. Linnemann C, et al. (2015) High-throughput epitope discovery reveals frequent rec-ognition of neo-antigens by CD4+ T cells in human melanoma. Nat Med 21:81–85.

48. Schenkel JM, Masopust D (2014) Tissue-resident memory T cells. Immunity 41:886–897.

49. Park CO, Kupper TS (2015) The emerging role of resident memory T cells in protectiveimmunity and inflammatory disease. Nat Med 21:688–697.

50. Li H, Ye C, Ji G, Han J (2012) Determinants of public T cell responses. Cell Res 22:33–42.51. Venturi V, Price DA, Douek DC, Davenport MP (2008) The molecular basis for public

T-cell responses? Nat Rev Immunol 8:231–238.52. Callan CG, Jr, Mora T, Walczak AM (2017) Repertoire sequencing and the statistical

ensemble approach to adaptive immunity. Curr Opin Syst Biol 1:44–47.53. Murugan A, Mora T, Walczak AM, Callan CG, Jr (2012) Statistical inference of the

generation probability of T-cell receptors from sequence repertoires. Proc Natl AcadSci USA 109:16161–16166.

54. Carlson CS, et al. (2013) Using synthetic templates to design an unbiased multiplexPCR assay. Nat Commun 4:2680.

Beausang et al. PNAS | Published online November 14, 2017 | E10417

IMMUNOLO

GYAND

INFLAMMATION

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

11, 2

020